My chosen corpus contains songs from playlists found on Spotify titled or closely resembling ‘autumn’, and in this portfolio I will attempt to find out what makes an ‘autumn’ song an ‘autumn’ song. I chose this corpus because a season is not universally associated with (a particular type of) music, and yet there are many playlists about specific seasons. What makes autumn especially so interesting is that despite it being the most disliked season of the four (at least in my experience), it has a lot of playlists dedicated to it, many of which do not contain sad music (as one would expect in a playlist about a least favourite season). There are also no direct reasons to create a playlist for this season, like going on a vacation could be a reason to create a summer playlist, or Christmas to create a winter playlist. The playlists incorporated in the corpus are bound to differ at least a little, since different people from different backgrounds who know and like different songs created them, but significant trends should be discernible. I decided not to include any playlists titled ‘fall’ or similar, due to the other meaning of the word ‘fall’, and ‘autumn’ is a much more specific word. How representative this corpus is, is difficult to say. Playlists like these also exist on other music streaming platforms like YouTube, and these can either be incredibly similar or very different, so it is difficult to judge the representativeness. I have tried my best to make this corpus as representative as possible of the playlists that can be found on Spotify, and I think I succeeded in that. Some very typical songs for this corpus include The Night We Met by Lord Huron, Cardigan by Taylor Swift, I Wanna Be Yours by the Arctic Monkeys, Yellow by Coldplay and Sweater Weather by The Neighbourhood. These are typical because they appear in several of the playlists I used and looked at for my corpus, indicating that all of the creators of the playlists thought of the same song, so the song must really fit the criteria.
This page shows a couple scatterplots, depicting several track-level features. This can provide a global overview of what exactly makes an ‘autumn’ song an autumn song.
The first graph depicts the valence on the x-axis and the danceability on the y-axis. It shows that the happier the song (which can be deducted from the valence, with 0 being the most negative sounding song and 1 being the most positive sounding song), the higher the danceability will be, which was to be expected. However, there is also a big cluster of songs that are classified as having a low valence that still score somewhat high on the danceability scale. Therefore, one of the assumptions that can be made is that some ‘autumn’ songs can be classified as negative sounding with still some level of danceability. In general, it can be said that autumn songs are more negative sounding songs with a low to medium level of danceability.
The second graph plots the instrumentalness against the speechiness. It is clear that the majority of the songs score very low on both scales, which could mean that there is a balance between music and the vocals of this corpus. Another observation that can be made from this graph is that there are also some tracks which score high on the instrumentalness scale, indicating that some songs contain little to no vocals. There is only one song that scores above 0,4, which most likely means it contains both music and spoken words.
The third graph visualises the acousticness on the x-axis and the liveness on the y-axis. It shows that this corpus contains a lot of acoustic songs, but just as many songs that are not acoustic. There are not many live performances in this corpus, but based on the ones that do exist, there is no correlation visible between liveness and acousticness. This is slightly surprising, since many artists perform their non acoustic songs acoustically during a live performance.
The fourth graph depicts the tempo on the x-axis and the energy on the y-axis. There is no correlation between the two visible, but it does show that an ‘autumn’ song cannot be defined by its tempo or energy. Globally speaking, the energy of these songs tend to be just below the halfway point (0,5), and the tempo tends to be in the range of 75-175 BPM.
The last graph on this page is a histogram which shows the loudness of all the songs included in this corpus. According to Spotify, the loudness usually ranges from -60 to 0. In my corpus, it ranges from -25 to 0, indicating that on average, the songs included in this corpus are louder than other songs on Spotify.
As stated in the introduction, one of the typical songs of this corpus is ‘The night we met’ by Lord Huron. The first plot is a cepstrogram depicting that song. The summarisation level of this graph is bars, since that provided the clearest results. There is a clear pattern visible, namely that the concentration is much higher in the lower bands, and much lower in the hihger bands.
The second plot is a chromagram of the same song. It shows that this song is mostly in A, but there are also short burst of energy in the D band across the song. The only other two significant engough to mention are F# and C#.
To compare two versions of a different song, both of which are included in this corpus, this page contains a graph comparing two versions of ‘willow’ by Taylor Swift, both of which are performed by Taylor Swift herself. It shows that there is a certain level of similarity between the songs, and the similarity seems to increase towards the end of the song. This could be because the end of the song seems to focus a bit more on vocals than earlier parts of the song, and the vocals do sound more or less the same in both versions of the song.
The first graph on this page is a histogram depicting how many songs are in which key in my corpus. It shows that C, G and A are the most prevalent keys for songs in this corpus.
Another song that is typical for my corpus is ‘Landslide’ by Fleetwood Mac, of which a keygram is visible here, showing which key every section of the song is in. The intro and bridge are clearly visible as two columns at the beginning of the song and just before the 100 seconds mark. In these sections, the key is a little less obvious. Looking at the entire keygram, however, it becomes clear that the key of this song is G minor, although B♭ major, D minor and F major are also clearly present in certain parts of the song.
For a general overview, the first graph on this page is a histogram of the tempi of the songs in this corpus. It shows that 120 BPM is by far the most common tempo in this corpus. This aligns with the readings from week 11, stating that 120 BPM is the average tempo for songs.
One of the artists featured most in my corpus is Hozier, and one of his songs that is typical for this corpus is ‘Like Real People Do’, which is visualised here in the form of a tempogram. It shows that the tempo, which is around 140 BPM, remains fairly constant throughout the song.
This corpus contains a lot of songs, which makes this visualisation a bit unclear. I also had to remove several songs from the corpus, because some songs had duplicate titles. I decided which ones to remove and which ones to keep based on which artist was not included in my corpus anywhere else, and if the song was a cover, which of the versions is the original.
From all of these graphs, a few conclusions can be made. Firstly, ‘autumn’ songs can often be classified as having a slightly negative sound and having a reasonable level of danceability. There is also often a good balance between instrumentals and vocals, and there are little live performances. Compared to other songs, ‘autumn’ songs can be considered loud. These sort of songs can be either acoustic or not. No conclusions could be made regarding tempo and energy, they seemed to have no influence on this genre. The most common keys in ‘autumn’ songs are C, G and A. The average tempo for this corpus was 120 BPM, which is about the same average as songs in other genres. If the tempogram in this portfolio is representative, then ‘autumn’ songs have a steady tempo throughout the entire song. Another observation that can be made from looking at the corpus is that it does not include songs from just on decade, but it includes both recent songs and songs from the 1970s. As far as I know, Spotify has no feature to identify when a song was released, but it could be interesting to see if there are certain patterns to be discovered in future research.
I have also discovered that, despite not liking autumn, I do like the playlists made about it, and will most likely continue to listen to these playlists, and maybe even compare them to playlists created on YouTube.
This portfolio has attempted to investigate what qualities a song that people can include in an ‘autumn’ playlist on Spotify has, and it provided several useful conclusions. There is also a lot of further research that can be done in this area, for example comparing these playlist to playlists made for other seasons, and looking at the similarities and differences.